Skip to content

Boltz-2

Summary

Boltz-2 is an open-source structural biology foundation model that extends the AlphaFold-3-style co-folding design of Boltz-1 with an integrated binding-affinity prediction module. It is reported to be the first AI model to approach the accuracy of physics-based free-energy perturbation (FEP) methods for protein-ligand binding affinity. It achieves this while running at least 1,000x faster than FEP.[1]

Architecture

Boltz-2's structure trunk expands the PairFormer stack from 48 to 64 layers, and uses bfloat16 mixed-precision training plus trifast triangle-attention kernels to make training on 768-token crops (up from 512 in Boltz-1) tractable.[1] Its confidence module is a lighter 8-PairFormer-layer head (vs. AlphaFold 3's 4) built on the final pair-token representation and predicted coordinates, replacing Boltz-1's much larger confidence trunk.[1]

Three controllability features are new relative to Boltz-1: method conditioning (specify X-ray, NMR, or MD as the target modality), template conditioning and steering (including native support for multimeric templates), and contact/pocket conditioning (user-specified distance constraints).[1] Physical-plausibility "Boltz-steering" potentials — previously an optional post-hoc add-on (Boltz-1x) — are integrated directly into inference (Boltz-2x), correcting steric clashes, incorrect stereochemistry, and non-planar aromatic rings without degrading geometric accuracy.[1]

The affinity module regresses log₁₀(IC₅₀/µM) together with a binder-probability score, trained on Ki/Kd/IC50/XC50/EC50/AC50-type measurements.[1]

Performance

  • Structure prediction: on a 2024–2025 PDB evaluation set disjoint from training data, Boltz-2 matches or moderately improves on Boltz-1 (largest gains in RNA and DNA-protein complexes, attributed to expanded distillation data), edges out Chai-1 and ProteinX, but trails AlphaFold 3 overall.[1]
  • Binding affinity vs. FEP: on a focused 4-target FEP+ subset (with training filtered to exclude ≥90% sequence identity to any FEP+ protein), Boltz-2 achieves an average Pearson correlation of 0.66, outperforming other inexpensive physical and ML baselines and approaching full free-energy-perturbation accuracy at >1,000x the speed; it also approaches OpenFE's performance on the full 876-measurement OpenFE benchmark.[1]
  • CASP16 affinity challenge (140 protein-ligand pairs, 2 targets): run out-of-the-box with no fine-tuning, Boltz-2 outperforms all top-ranking competition entries, which had access to weeks of ad-hoc tool development.[1]
  • Retrospective hit discovery (MF-PCBA): nearly doubles the average precision of prior ML approaches and of ipTM/docking baselines, reaching an enrichment factor of 18.4 at the top 0.5% threshold.[1]
  • Private/internal assays: on eight blinded Recursion medicinal-chemistry assays, Boltz-2 exceeds other ML baselines but achieves Pearson R>0.55 on only 3 of 8 — comparable variability to FEP itself, which is known to struggle on some target classes (e.g. GPCRs) without bespoke input preparation.[1]

Prospective Virtual Screening

The authors validated a generative screening workflow on the kinase target TYK2: Boltz-2 scored a fixed library and, separately, molecules generated by SynFlowNet (a GFlowNet sampler over Enamine's 76-billion-compound REAL space), with top candidates checked against Boltz-ABFE (an absolute-FEP pipeline; Pearson R=0.95, MAE=0.42 kcal/mol vs. experiment on a benchmark set). The SynFlowNet stream produced higher average predicted affinity using roughly a quarter of the model evaluations of the fixed-library screen (117k vs. 460k), and generated compounds were not close analogues (by Tanimoto similarity) of known PDB-derived TYK2 binders.[1]

Availability

Weights, inference, and training code are released under a permissive open license at github.com/jwohlwend/boltz. This is a preprint, not yet peer-reviewed.

See Also

  • AlphaFold 3 and Protenix — co-folding architectures Boltz-2 is benchmarked against.
  • Chai-1 — a contemporary co-folding foundation model without a binding-affinity module.
  • DrugCLIP — a contrastive-retrieval alternative for genome-wide virtual screening rather than pose/affinity co-prediction.

Citations

[1] Passaro, Corso, Wohlwend et al. (2025), "Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction", bioRxiv Supports: architecture, all benchmark and screening figures above. Location: Sections 2 (Architecture), 5 (Results: structure, affinity, virtual screening).